Deep frequency-domain adaptive filtering convolution for intelligent fault diagnosis of rotating machinery

Deep neural networks, represented by convolutional neural networks (CNNs), have been widely used in fault diagnosis. In recent years, frequency-domain modeling can complement spatial-domain convolution by capturing global spectral structures. In this work, a global feature learning module based on the image Fourier transform, termed deep frequency-domain adaptive filtering (DFAF), is proposed. We first observe that the time–frequency images of mechanical vibration signals tend to exhibit low-frequency components along the time direction, leading to a vertical concentration of spectral energy. The corresponding theoretical basis is given. Then, a frequency filter with a learnable bandwidth is designed to generate a soft, smooth mask that encourages the model to focus on frequency components beneficial to noise robustness diagnosis. The proposed DFAF module is simple, lightweight, and plug-and-play. It provides a nonlocal receptive field and enables multiscale information without substantially increasing the parameter count. By inserting the DFAF module into a CNN, we construct a spatial-frequency fusion convolutional (SFConv) network and evaluate it on two rotating machinery fault datasets. At signal-to-noise ratio = −10 dB, SFConv exceeds the best comparison methods by 6.68 and 3.26 percentage points.